← Back to all work

Multi-agent automation infrastructure

The problem

Running multiple specialized LLM agents reliably for ongoing personal automation: messaging integration, model provider routing, cost control, and resilience against provider changes.

What I built

Self-hosted infrastructure on Hetzner Cloud running an open-source agent framework with multiple Telegram-connected agents, primary model routing through one provider with fallbacks to others, and migration tooling between providers as the model landscape shifted.

Stack

Linux server on Hetzner Cloud (CX33), open-source agent framework, multiple LLM providers (Anthropic, MiniMax, Gemini), Telegram integration.

What I learned

• Context-window strategy is the silent variable when migrating providers. Different models charge, truncate, and summarize context in different ways — agent behavior changes long before the model "fails," so migration needs eval coverage, not just smoke tests.
• The economics of self-hosting agent infrastructure are nuanced. A small managed cloud node beats a managed agent service on cost and control once you have steady workloads, but only if you're willing to own runtime, secrets, and upgrades. Below a usage threshold, managed services win.

Related work

• cravingtoolkit.com — RAG and AI content pipeline — production AI in a YMYL domain.
• Order Management Systems at dwpbank — production Spring Boot, Kafka, Angular at financial-services scale.